SRTM与ASTER加权融合的机器学习方法

A Weighted Fusion Method of SRTM and ASTER Based on Machine Learning

为了克服SRTM和ASTER各自缺陷,充分结合二者优势得到更高质量的DEM,提出了一种基于神经网络模型的加权融合方法。首先,统一两种DEM坐标系和高程基准;其次,借助后向传播神经网络分别建立SRTM与ASTER高程误差和地形因子的非线性关系模型;然后,利用此模型估计各DEM的误差分布;最后,根据该误差计算SRTM和ASTER融合权重,并实现SRTM和ASTER加权融合。以董志塬为研究区进行分析。结果表明:融合后DEM精度有明显提高,相比于原始SRTM和ASTER,平均绝对误差分别降低了1.29 m和3.66 m,中误差分别降低了0.40 m和3.16 m,标准差分别降低了0.79 m和2.07 m;各地形因子对DEM高程精度的影响在融合之后均得到降低。以美国爱达荷州北部某区域为验证区,实验结果与研究区相似。

In order to overcome the shortcomings of SRTM and ASTER and fully combine their advantages to produce high-quality DEM,this paper develops a weighted fusion method based on neural network prediction model.Firstly,the coordinate systems and elevation datum of two public DEMs are unified firstly.Then,nonlinear relationship models between SRTM and ASTER errors and all terrain factors are constructed by back propagation neural network,respectively.Next,the distributions of DEMs errors are fitted by the models.Finally,the two DEMs are fused based on the fusion weights that are calculated according to the elevation errors.The results of study area indicate that the precision of the fused DEM is improved,with the mean absolute error decreased by 1.29 m and 3.66 m,the elevation mean error 0.40 m and 3.16 m lower and the elevation standard deviation reduced by 0.79 m and 2.07 m,respectively,compared with the original SRTM and ASTER.The influence of terrain factors on DEM elevation accuracy is reduced after data fusion.Selecting a region in northern Idaho as the verification area,the experimental results are similar to those in the study area.